How I stopped my AI agents from getting dumber after 10 turns

TL;DR: Built an open-source context engineering library. 5/5 decision recall vs 4/5 naive on Claude Opus, 52% token reduction, 2.7x information density, 100% cache hit rate. Apache 2.0. Link at the bottom if you want to skip the story. The Journey (skip if you don't like storytelling) Let me preface this — I am in no way an authority on AI agent design (this is literally my first serious project in this space), just a guy with a deep hyperfocus that found an issue and went after trying to sort it out. I've been building AI agents locally for the past few months — personal assistant, task runner, coding helper, the whole shebang. This started when the whole open-source agent hype exploded. I got excited, tried one of the popular ones, and before I even sent a single message the context was already sitting at 15K+ tokens. Fresh install, no conversation, 15K gone. Then I tried another that looked promising but needed serious elbow grease to set up and was way more complex than what I need